Abstract

Often considered as the ground truth, wireline logs that are acquired along wells and are sensitive to formation properties serve as the foundation for reservoir characterization workflows. Effectively utilizing these well logs necessitate integrated geological, petrophysical, geomechanical expertise. The log availability and data quality, in general, varies along the reservoir. We propose a machine learning workflow to first improve wellbore log quality, then propagate markers as well as petrophysical and geomechanical property interpretation from a few high tier wells to the entire field where log availability might be restricted. This DL-based approach significantly expedites the reservoir characterization process while mitigating the uncertainties associated with human bias.

The workflow incorporates several DL steps. Initially, a log quality control (QC) process is applied to generate a complete set of logs, with corrections made for outlier samples. We then manually interpret key markers on sparse wells, which serve as the basis for training an DL model for marker propagation. This model predicts major markers on the remaining wells, and a marker QC step is implemented to further refine the results. The identified markers define the major formations in the study area. For each formation, DL-based property interpretation models are trained to predict petrophysical properties such as water saturation, total organic carbon, and total porosity. Additionally, a separate property interpretation model is trained for each formation for geomechanical properties.

We employ data from two unconventional plays in North America to illustrate the proposed workflow. In both instances, we note that the logs, post the QC process, exhibit greater consistency, and more accurately reflect the true formation properties compared to the original logs. This QC process proves especially pivotal for subsequent DL-based marker propagation and property interpretation, due to the importance of high-quality training data. The marker propagation model achieves nearly 100% accuracy for major markers within a 2-foot tolerance. The resultant petrophysical and geomechanical properties within each formation provide complete coverage of 1D properties, that are ready for 3D subsurface model building (both static and dynamic) through integration with seismic data.

The proposed approach facilitates a unified workflow for wellbore logs data starting from data cleaning, gap filling, to marker picking and marker QC, to log interpretation. Such a workflow can be easily plugged-and-played into a variety of workflows that fall under the umbrella of formation evaluation such as petrophysics, geomechanics, acoustics etc. The steps in the unified workflow are implemented using a wide variety of DL models ranging from classical ML-based approaches to autoencoder based neural networks to the more powerful Transformer-based approaches that are very effective at modelling the sequential nature of the wellbore logs.

This content is only available via PDF.
You can access this article if you purchase or spend a download.